Bricken and colleagues: sparse-autoencoder language-model feature decomposition

What changed

Extracted sparse learned features from a one-layer Transformer's activations and evaluated their interpretability, interventions and feature-splitting behavior.

Scientific scope

Extracted sparse learned features from a one-layer Transformer's activations and evaluated their interpretability, interventions and feature-splitting behavior. Does not invent dictionary learning or sparse autoencoders. Interpretable features in a small model do not amount to complete model understanding or a safety guarantee.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Does not invent dictionary learning or sparse autoencoders. Interpretable features in a small model do not amount to complete model understanding or a safety guarantee.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction